| Message ID | 20200522153901.133375-12-hare@suse.de (mailing list archive) |
|---|---|
| State | Superseded, archived |
| Delegated to: | Mike Snitzer |
| Headers | show |
| Series | dm-zoned: multi-device support | expand |
On 2020/05/23 0:39, Hannes Reinecke wrote: > When reclaiming zones we should arbitrate between the zoned > devices to get a better throughput. So implement a simple > round-robin load balancer between the zoned devices. > > Signed-off-by: Hannes Reinecke <hare@suse.de> > --- > drivers/md/dm-zoned-metadata.c | 8 +++++++- > 1 file changed, 7 insertions(+), 1 deletion(-) > > diff --git a/drivers/md/dm-zoned-metadata.c b/drivers/md/dm-zoned-metadata.c > index 87784e7785bc..25dcad2a565f 100644 > --- a/drivers/md/dm-zoned-metadata.c > +++ b/drivers/md/dm-zoned-metadata.c > @@ -171,6 +171,8 @@ struct dmz_metadata { > unsigned int nr_reserved_seq; > unsigned int nr_chunks; > > + unsigned int last_alloc_idx; > + > /* Zone information array */ > struct xarray zones; > > @@ -2178,7 +2180,7 @@ struct dm_zone *dmz_alloc_zone(struct dmz_metadata *zmd, unsigned long flags) > { > struct list_head *list; > struct dm_zone *zone; > - unsigned int dev_idx = 0; > + unsigned int dev_idx = zmd->last_alloc_idx; > > again: > if (flags & DMZ_ALLOC_CACHE) > @@ -2214,6 +2216,9 @@ struct dm_zone *dmz_alloc_zone(struct dmz_metadata *zmd, unsigned long flags) > zone = list_first_entry(list, struct dm_zone, link); > list_del_init(&zone->link); > > + if (!(flags & DMZ_ALLOC_CACHE)) > + zmd->last_alloc_idx = (dev_idx + 1) % zmd->nr_devs; > + > if (dmz_is_cache(zone)) > atomic_dec(&zmd->unmap_nr_cache); > else if (dmz_is_rnd(zone)) > @@ -2839,6 +2844,7 @@ int dmz_ctr_metadata(struct dmz_dev *dev, int num_dev, > zmd->dev = dev; > zmd->nr_devs = num_dev; > zmd->mblk_rbtree = RB_ROOT; > + zmd->last_alloc_idx = 0; > init_rwsem(&zmd->mblk_sem); > mutex_init(&zmd->mblk_flush_lock); > spin_lock_init(&zmd->mblk_lock); > OK. So my comment on patch 8 is already addressed. Or at least partly... Where is last_alloc_idx actually used ? It looks like this only sets last_alloc_idx but do not use that value on entry to dmz_alloc_zone() to allocate the zone.
On 5/25/20 4:42 AM, Damien Le Moal wrote: > On 2020/05/23 0:39, Hannes Reinecke wrote: >> When reclaiming zones we should arbitrate between the zoned >> devices to get a better throughput. So implement a simple >> round-robin load balancer between the zoned devices. >> >> Signed-off-by: Hannes Reinecke <hare@suse.de> >> --- >> drivers/md/dm-zoned-metadata.c | 8 +++++++- >> 1 file changed, 7 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/md/dm-zoned-metadata.c b/drivers/md/dm-zoned-metadata.c >> index 87784e7785bc..25dcad2a565f 100644 >> --- a/drivers/md/dm-zoned-metadata.c >> +++ b/drivers/md/dm-zoned-metadata.c >> @@ -171,6 +171,8 @@ struct dmz_metadata { >> unsigned int nr_reserved_seq; >> unsigned int nr_chunks; >> >> + unsigned int last_alloc_idx; >> + >> /* Zone information array */ >> struct xarray zones; >> >> @@ -2178,7 +2180,7 @@ struct dm_zone *dmz_alloc_zone(struct dmz_metadata *zmd, unsigned long flags) >> { >> struct list_head *list; >> struct dm_zone *zone; >> - unsigned int dev_idx = 0; >> + unsigned int dev_idx = zmd->last_alloc_idx; >> >> again: >> if (flags & DMZ_ALLOC_CACHE) >> @@ -2214,6 +2216,9 @@ struct dm_zone *dmz_alloc_zone(struct dmz_metadata *zmd, unsigned long flags) >> zone = list_first_entry(list, struct dm_zone, link); >> list_del_init(&zone->link); >> >> + if (!(flags & DMZ_ALLOC_CACHE)) >> + zmd->last_alloc_idx = (dev_idx + 1) % zmd->nr_devs; >> + >> if (dmz_is_cache(zone)) >> atomic_dec(&zmd->unmap_nr_cache); >> else if (dmz_is_rnd(zone)) >> @@ -2839,6 +2844,7 @@ int dmz_ctr_metadata(struct dmz_dev *dev, int num_dev, >> zmd->dev = dev; >> zmd->nr_devs = num_dev; >> zmd->mblk_rbtree = RB_ROOT; >> + zmd->last_alloc_idx = 0; >> init_rwsem(&zmd->mblk_sem); >> mutex_init(&zmd->mblk_flush_lock); >> spin_lock_init(&zmd->mblk_lock); >> > > > OK. So my comment on patch 8 is already addressed. Or at least partly... Where > is last_alloc_idx actually used ? It looks like this only sets last_alloc_idx > but do not use that value on entry to dmz_alloc_zone() to allocate the zone. > Aw, fsck. Something went astray when generating the patches. Will be fixing it up for the next round. Cheers, Hannes
diff --git a/drivers/md/dm-zoned-metadata.c b/drivers/md/dm-zoned-metadata.c index 87784e7785bc..25dcad2a565f 100644 --- a/drivers/md/dm-zoned-metadata.c +++ b/drivers/md/dm-zoned-metadata.c @@ -171,6 +171,8 @@ struct dmz_metadata { unsigned int nr_reserved_seq; unsigned int nr_chunks; + unsigned int last_alloc_idx; + /* Zone information array */ struct xarray zones; @@ -2178,7 +2180,7 @@ struct dm_zone *dmz_alloc_zone(struct dmz_metadata *zmd, unsigned long flags) { struct list_head *list; struct dm_zone *zone; - unsigned int dev_idx = 0; + unsigned int dev_idx = zmd->last_alloc_idx; again: if (flags & DMZ_ALLOC_CACHE) @@ -2214,6 +2216,9 @@ struct dm_zone *dmz_alloc_zone(struct dmz_metadata *zmd, unsigned long flags) zone = list_first_entry(list, struct dm_zone, link); list_del_init(&zone->link); + if (!(flags & DMZ_ALLOC_CACHE)) + zmd->last_alloc_idx = (dev_idx + 1) % zmd->nr_devs; + if (dmz_is_cache(zone)) atomic_dec(&zmd->unmap_nr_cache); else if (dmz_is_rnd(zone)) @@ -2839,6 +2844,7 @@ int dmz_ctr_metadata(struct dmz_dev *dev, int num_dev, zmd->dev = dev; zmd->nr_devs = num_dev; zmd->mblk_rbtree = RB_ROOT; + zmd->last_alloc_idx = 0; init_rwsem(&zmd->mblk_sem); mutex_init(&zmd->mblk_flush_lock); spin_lock_init(&zmd->mblk_lock);
When reclaiming zones we should arbitrate between the zoned devices to get a better throughput. So implement a simple round-robin load balancer between the zoned devices. Signed-off-by: Hannes Reinecke <hare@suse.de> --- drivers/md/dm-zoned-metadata.c | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-)